Generally, various rock types—with their own specific rock-forming minerals—have their own reflectance signatures, and can thus be discriminated based on their spectral characteristics using different spectral wavelengths of optical remote-sensing images (Kang et al., 2001; Corumluoglu et al., 2015; Hassan and Ramadan, 2015). Satellite images have been used to map the Earth for decades (Dou et al., 2015a; Dou et al., 2015b; Lillesand et al., 2015; Moghtaderi et al., 2022), and multispectral sensors are widely used for lithological discrimination in areas where rock units are exposed (Haselwimmer et al., 2010; Nair and Mathew, 2012; Arivazhagan and Anbazhagan, 2017). In comparison to hyperspectral imagery, multispectral imagery is more advantageous in terms of classifying major lithologies (e.g., basalt, granite, and rhyolite) due to its high quality, high spatial resolution, low cost, and high accessibility (Ge et al., 2018). Nevertheless, hyperspectral imagery is more suitable for the finer classifications of rocks (e.g., tholeiitic basalt, alkali basalt, etc.).

Landsat Thematic Mapper™, the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), and Sentinel-2 are the most widely used multispectral sensors for the discrimination of mixed rock units. Of the three, ASTER and Sentinel-2 are regarded to be more suitable due to their increased availability of infrared bands and better spatial resolution than the Landsat Thematic Mapper (Gomez et al., 2005; Perry and Kruse, 2010; Ehlers and Klonus, 2014). In this study, we use the geological survey of the basalt distribution by the Sichuan Geological Survey in 2005 as prior knowledge to assess the remote-sensing identifications. Further, the Sentinel-2 sensor began collecting images after 2015; therefore, we did not use Sentinel-2 imageries here considering that the vegetation coverage changed significantly in the geological-survey area from 2005 to 2015. Instead, we used ASTER and Landsat-5 TM images collected from 2003 to 2007 and in 2004. Three machine learning methods, including k-nearest neighbors (KNN), maximum likelihood classification (MLC), and support vertical machine (SVM), were compared to identify basalts based on both ASTER and Landsat-5 data. Therefore, the prediction results of six models, including the KNN prediction of basalts by ASTER images, SVM prediction by ASTER, MLC prediction by ASTER, KNN prediction by Landsat-5 images, SVM prediction by Landsat-5, and MLC prediction by Landsat-5, were compared based on the area of bare basalt in the geological-survey scope, to calculate the performance metrics (precision and accuracy). The geological-survey basalt is partly covered by vegetation that is thus hard to be identified by remote-sensing imageries, so we just used the bare-basalt distribution of the geological-survey, that is without vegetation coverage, to assess the identified results. The bare-basalt distribution is extracted through land-use classification for the geological-survey scope (bare land and vegetation).

Furthermore, the flood basalts are the signature feature of the Emeishan Large Igneous Province (LIP) in spite that there are also ultramafic and silicic volcanic rocks and layered mafic-ultramafic and silicic plutonic rocks exposed (Shellnutt, 2014). There were much substantial researches on ELIP in the past, and the reason why it is of particular interest is that it contains numerous world-class orthomagmatic Fe-Ti-V deposits and series of smaller economically important Ni-Cu-(OGE) sulphide deposits but also is contemporaneous with the Late Capitanian (∼260 Ma) mass extinction (Zhou et al., 2002; Zhang et al., 2006; Ganino and Arndt 2009). Numerous researches on ELIP in the past covered a wide scope of geology, paleomagnetism, geochronology, geochemistry, biostratigraphy and so on. But remote sensing technology was seldom used in the researches on ELIP flood basalts. This paper conducted the studies of remote-sensing identification on flood basalts in the Panxi part of ELIP.

We analyzed three supervised-classification algorithms and two kinds of remote-sensing imageries, as mentioned previously, to identify basalts using the training images. These training images are part of the Landsat-5 imagery data located at the sampling sites, as described in Figure 2, when identification uses Landsat-5 images; they are also part of ASTER imagery, when identification uses ASTER images. In total, six identification methods are used: the KNN prediction of basalts by ASTER images, SVM prediction by ASTER, MLC prediction by ASTER, KNN prediction of basalts by Landsat-5 images, SVM prediction by Landsat-5, and MLC prediction by Landsat-5. All identifications are carried out using ENVI 5.3.

When using MLC, the probability threshold is set at 99%. This threshold is a probability minimum for inclusion in a class, and the ENVI does not classify pixels with a value lower than this value. As a binary-classification method, we tried changing the probability threshold of MLC from 51% to 99.9%; it was found that a threshold below 99% led to anomalous data that deviated from prior-knowledge basalt areas.

The results of the six identification methods are shown in Figure 4. Generally, the identified locations of basalts are concentrated in the prior areas of geological-survey basalt, which indicates that the methods of identification used (including the selections of both remote-sensing images and algorithms) are reasonable. Furthermore, in past research, remote-sensing images for identifying the lithology exhibit poor performance with significant vegetation coverage (Grebby et al., 2014). Therefore, in our study, the samples used for training are chosen to be bare and without vegetation coverage; here, remote-sensing images simply identify basalts that are exposed and cannot detect basalts under vegetation. We overlap bare land (Figure 3) with the identified results (Figure 4), taking the area of geological-survey basalts as prior knowledge to calculate the precision and accuracy for the six identification methods. Figure 5A shows that how the three ASTER identification results fall into the prior area of basalts, and Figure 5B shows that how the three Landsat-5 identification results fall into the prior area of basalts. Assuming that the geological-survey basalt areas are correct, the performance metrics (including precision and accuracy) of the six identification methods are shown in Table 1.

As shown in Figure 5, most of the identified area falling into the geological-survey scope is on bare land, except a minor portion of the identified area from SVM by Landsat-5, which slightly overlaps with vegetation. It is inevitable that all methods exhibit certain errors, but the overall identification of basalts by remote-sensing images, with the training data of bare-basalt images, is still correct for bare land.

The three algorithms of KNN, MLC, and SVM are all object-oriented identifications; therefore, the identified basalts have many clusters with various sizes. Generally, the identified clusters by Landsat-5 show larger sizes than those by ASTER (Figure 4; Figure 5), possibly because Landsat-5 has a lower spatial resolution of 30 m for identification, than ASTER’s 15 m, which is better for our study area with the total area of 8328 km².

Moreover, it seems that the SVM algorithm performs most effectively on both ASTER and Landsat-5 imageries than the other two algorithms, because the total identified area of basalts by SVM is slightly more. KNN performed very well on Landsat-5; MLC performed poorest on both ASTER and Landsat-5 imageries (Figure 5 and Table 2), compared to SVM and KNN. The SVM algorithm is essentially a two-category classifier of deep learning and is possibly more robust than the other two algorithms of simple machine learning. When the probability threshold of MLC is set lower, it produces results that deviate from prior knowledge; when it is set higher, it produces a smaller area of prediction. The error emergence of singular solutions possibly occurs when solving the covariance matrix in the MLC algorithm. KNN performs intermediately between SVM and MLC, and predicts more area based on Landsat-5 than based on ASTER, possibly only because the spatial resolution of Landsat-5 TM imagery is 30 m, in contrast to the 15-m resolution of ASTER imagery.

Although ASTER is widely used for the discrimination of mixed rock units, Landsat-5 data has advantages in certain cases, especially when the total area for identification is large. However, ASTER data would be more advantageous for more detailed identification in smaller areas. Furthermore, for machine learning algorithms for the supervised classification within ENVI, SVM may perform relatively more robustly.

Generally, in this study, we use both ASTER and Landsat-5 data for discrimination of the mixed rocks by three supervised-classification algorithms, and in the training, small-size training images are utilized. As a result, a total of six models are constructed. The modeling results by training small-size images are generally in accordance with the trends of the prior area of geological-survey basalt distributions. This verifies the capability of remote-sensing imagery for lithological discrimination and the capability of the SVM algorithm for deep learning in supervised classification. However, there are still issues in terms of how remote-sensing imagery can identify rocks with high vegetation coverage. Under the constraints of present remote-sensing technology, the only possibility may be to develop algorithms of deep learning neural networks.

Besides, this work of the remote-sensing identification of flood basalts in Panxi part of Emeishan LIP is significant, which will promote us to launch a wider-scope identification across Emeishan LIP in the future. Remote-sensing identification of flood basalts all over the Emeishan LIP will not only help map the distribution of flood basalts, but also help explore the world-class Fe-Ti-V deposits and Ni-Cu-(OGE) sulphide deposits that are both closely related to flood basalt distributions.